Personalized Multitask Learning for Predicting Tomorrow's Mood, Stress, and Health

Abstract

While accurately predicting mood and wellbeing could have a number of important clinical benefits, traditional machine learning (ML) methods frequently yield low performance in this domain. We posit that this is because a one-size-fits-all machine learning model is inherently ill-suited to predicting outcomes like mood and stress, which vary greatly due to individual differences. Therefore, we employ Multitask Learning (MTL) techniques to train personalized ML models which are customized to the needs of each individual, but still leverage data from across the population. Three formulations of MTL are compared: i) MTL deep neural networks, which share several hidden layers but have final layers unique to each task; ii) Multi-task Multi-Kernel learning, which feeds information across tasks through kernel weights on feature types; and iii) a Hierarchical Bayesian model in which tasks share a common Dirichlet Process prior. We offer the code for this work in open source. These techniques are investigated in the context of predicting future mood, stress, and health using data collected from surveys, wearable sensors, smartphone logs, and the weather. Empirical results demonstrate that using MTL to account for individual differences provides large performance improvements over traditional machine learning methods and provides personalized, actionable insights.

Keywords: Deep Neural Networks; Hierarchical Bayesian Model; Mood Prediction; Multi-Kernel SVM; Multitask learning.

Fig. 1

A simplified version of the MTL-NN architecture. Clusters of related people receive specialized predictions from a portion of the network trained with only their data. Shared initial layers extract features relevant to all clusters.

Fig. 2

Distribution of self-report labels after discarding the middle 20%. Participants are listed on the x-axis, in order of their average self-report value for that label (each participant is one column). Almost all participants have data from both label classes.

Fig. 3

Features extracted for each detected, non-artifact SCR

Fig. 4

SMS frequency over four days

Fig. 5

Percent of participants sleeping, studying, in extracurricular activities, and exercising throughout the day.

Fig. 6

GMM fitted to location data from one participant. Black points are locations visited; the contours mark the probability distribution induced by the model.

Fig. 7

Accuracy for each type of model in the STL, MTL-moods, and MTL-people approaches. Note that the accuracy significantly (* = p < 0.05) improves when using multi-tasking over people for each label and for each machine learning method tested.

Fig. 8

MTMKL kernel modality weights, reflecting which feature type is most important to the classifier for each task. The ν parameter controls how heavily the task weights are regularized to be similar, and was set by the hyperparameter search.

Fig. 9

Resulting soft clustering (Φ) when predicting the different labels (mood, stress, and health). Each row shows one of the 104 participant’s degree of membership in each cluster. We note that there were 4,3, and 17 clusters needed in predicting happiness, stress, and health, respectively.

Fig. 10

Distribution of HBLR weights on the total number of screen on events (5pm-midnight) feature for each cluster when predicting tomorrow’s mood

Fig. 11

Example of different weight distributions induced by the soft clustering for 3 different participants in the mood prediction. Participant 3 is almost exclusively in cluster 1, participant 5 is has membership in clusters 0, 1, and 2, and participant 31 is almost exclusively in cluster 2.

Fig. 12

Mean feature weights for mood clusters in HBLR model. We note that the positive label is “Happy” so features with positive (negative) mean weights contribute to being more happy (sad) tomorrow.

Fig. 13

Mean feature weights for stress clusters in HBLR model. We note that the positive label is “Calm” so features with positive (negative) mean weights contribute to being more calm (stressed) tomorrow.